Projects

An agentic document Q&A system built on a 13-node LangGraph DAG. Routes each query through intent classification, multi-source retrieval (vector search, table extraction, web), and iterative refinement — with human-in-the-loop review before answering. Responses include source citations, chain-of-thought reasoning, and per-session cost tracking.

Tech Stack: Python, LangGraph, LangChain, LiteLLM, Qdrant, Redis, Postgres, Streamlit, Langfuse, Docker

A graph-based evaluation toolkit for LLM and RAG systems with repeatable quality checks, upfront cost visibility, cache for reusability and clean per-case outputs for analysis. Metric suport: Grounding, Relevance, RedTeam, Geval, Reference-based.

Tech Stack: Python, llm-as-judge, llvm-eval, llm-eval, cache

Perception and control modules for autonomous vehicles built on the Udacity SDC curriculum. Covers lane detection, traffic sign classification, behavioral cloning, LIDAR/RADAR sensor fusion via Extended Kalman Filter, jerk-minimizing path planning, and a PID controller for steering and throttle.

Tech Stack: Python, C++, OpenCV, TensorFlow, Keras, Eigen

Core Deep RL algorithms implemented across Unity ML-Agents environments. Covers DQN and Double-DQN for discrete action spaces, REINFORCE for Atari Pong, DDPG for continuous robotic arm control, and Multi-Agent DDPG for a collaborative/competitive multi-agent setting.

Tech Stack: Python, PyTorch, Unity ML-Agents, NumPy

Object detection and semantic segmentation using Faster-RCNN and Mask-RCNN on GPU. Benchmarks near-real-time two-stage detectors against YOLO's single-pass real-time approach, evaluating detection precision, segmentation quality, and inference speed across model configurations.

Tech Stack: Python, Scikit-learn, OpenCV, TensorFlow (GPU), Keras

Classifies real-estate as land or house from address strings and property imagery. Fuses building boundaries from OpenStreetMap with satellite and street-view images from Google Maps, using deep learning models, semantic segmentation, and ensembles to produce best-accuracy predictions.

Tech Stack: Python, Scikit-learn, OpenCV, GeoPandas, Shapely, TensorFlow (GPU), Keras, Azure Cloud stack

Face recognition built from scratch in TensorFlow following the FaceNet paper. Applies transfer learning on GoogLeNet (NN4 small) — Haar Cascades first extract face regions from an image, then learned facial embeddings are used to classify identities.

Tech Stack: Python, TensorFlow, OpenCV

Identifies credit card fraud using Bagging, Boosting, Deep Neural Nets, Autoencoders, and Bayesian methods with MCMC sampling. Models are evaluated on the Kaggle dataset and tuned for high recall on an imbalanced fraud detection problem.

Tech Stack: Python, R, TensorFlow, XGBoost (Python API)

Evaluates Deep Learning, image processing, and classical ML models for object classification on the CIFAR-10 dataset. Uses a controlled two-class subset to benchmark CNN architectures against traditional methods.

Tech Stack: Python, Scikit-learn, OpenCV, TensorFlow

Predicts early hospital re-admission of diabetic patients within 30 days using the UCI dataset. Analyzes contributing clinical factors and builds a classification pipeline with PySpark MLlib for scalable distributed processing.

Tech Stack: Spark, PySpark (MLlib), Python

Evaluates classification and regression algorithms for urban crime prediction. Determines whether a locality is high or low crime (classification) and forecasts the crime rate as a continuous value (regression).

Tech Stack: Python, Scikit-learn

Detects and extracts vehicle license plates from images using contour-based segmentation and binary classification. Trained on manually cropped plates bootstrapped to a larger dataset, classifying each contour candidate as a license plate or background.

Tech Stack: Python, Scikit-learn, OpenCV